app/utils/predict_afteruse.py

import torch
from transformers import BertTokenizer, BertForSequenceClassification
import numpy as np
import os

HF_TOKEN = os.getenv("HF_TOKEN")

# Load the pre-trained model and tokenizer
tokenizer = BertTokenizer.from_pretrained("Maulidaaa/bert-safe-multilabel", token=HF_TOKEN)
model = BertForSequenceClassification.from_pretrained("Maulidaaa/bert-safe-multilabel", token=HF_TOKEN)

model.eval()

# Efek after use (harus sesuai saat training)
afteruse_labels = [
    "acne fighting", "acne trigger", "anti aging", "brightening", "moisturizing",
    "redness reducing", "skin texture", "soothing", "unknown", "whitening"
]

# Deskripsi efek after use dalam bahasa Indonesia
afteruse_descriptions_id = {
    "acne fighting": "membantu melawan jerawat",
    "acne trigger": "dapat memicu jerawat",
    "anti aging": "mengurangi tanda-tanda penuaan",
    "brightening": "mencerahkan kulit",
    "moisturizing": "melembapkan kulit",
    "redness reducing": "mengurangi kemerahan pada kulit",
    "skin texture": "memperbaiki tekstur kulit",
    "soothing": "menenangkan kulit",
    "unknown": "memiliki efek yang belum diketahui",
    "whitening": "memutihkan kulit"
}

def predict_after_use(input_ingredients):
    if not input_ingredients:
        return []

    # Tokenisasi input
    inputs = tokenizer(input_ingredients, return_tensors="pt", truncation=True, padding=True, max_length=512)

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.sigmoid(logits)

    # Konversi probabilitas ke numpy dan ratakan
    probs = probs.squeeze().cpu().numpy()
    if probs.ndim == 0:
        probs = [probs]  # Jika scalar, ubah jadi list

    print(f"[DEBUG] Model output shape: {logits.shape}")
    print(f"[DEBUG] Jumlah output model: {len(probs)} | Jumlah label: {len(afteruse_labels)}")

    # Antisipasi mismatch jumlah output vs label
    min_len = min(len(probs), len(afteruse_labels))
    predicted_labels = [
        afteruse_labels[i]
        for i in range(min_len)
        if probs[i] > 0.5
    ]

    return predicted_labels

def generate_afteruse_sentence_id(predicted_labels):
    if not predicted_labels:
        return "Tidak ada efek yang terdeteksi berdasarkan bahan yang diberikan."

    descriptions = [afteruse_descriptions_id.get(label, label) for label in predicted_labels]

    if len(descriptions) == 1:
        return f"Produk ini {descriptions[0]}."
    elif len(descriptions) == 2:
        return f"Produk ini {descriptions[0]} dan {descriptions[1]}."
    else:
        return f"Produk ini {', '.join(descriptions[:-1])}, dan {descriptions[-1]}."

def predict_after_use_with_probs(input_ingredients):
    if not input_ingredients:
        return [], []

    inputs = tokenizer(input_ingredients, return_tensors="pt", truncation=True, padding=True, max_length=512)

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.sigmoid(logits)

    probs = probs.squeeze().cpu().numpy()
    if probs.ndim == 0:
        probs = [probs]

    min_len = min(len(probs), len(afteruse_labels))
    predicted_labels = [
        afteruse_labels[i]
        for i in range(min_len)
        if probs[i] > 0.5
    ]
    label_probs = {
        afteruse_labels[i]: float(probs[i])
        for i in range(min_len)
    }

    return predicted_labels, label_probs